CN110969626B - Method for extracting hippocampus of human brain nuclear magnetic resonance image based on 3D neural network - Google Patents
Method for extracting hippocampus of human brain nuclear magnetic resonance image based on 3D neural network Download PDFInfo
- Publication number
- CN110969626B CN110969626B CN201911179566.4A CN201911179566A CN110969626B CN 110969626 B CN110969626 B CN 110969626B CN 201911179566 A CN201911179566 A CN 201911179566A CN 110969626 B CN110969626 B CN 110969626B
- Authority
- CN
- China
- Prior art keywords
- hippocampus
- neural network
- convolution
- loss function
- module
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 210000001320 hippocampus Anatomy 0.000 title claims abstract description 90
- 210000004556 brain Anatomy 0.000 title claims abstract description 26
- 238000005481 NMR spectroscopy Methods 0.000 title claims abstract description 18
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 14
- 238000000034 method Methods 0.000 title claims description 29
- 230000011218 segmentation Effects 0.000 claims abstract description 61
- 238000003062 neural network model Methods 0.000 claims abstract description 28
- 238000007781 pre-processing Methods 0.000 claims abstract description 7
- 238000000605 extraction Methods 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 20
- 230000004913 activation Effects 0.000 claims description 12
- 238000010606 normalization Methods 0.000 claims description 9
- 238000011176 pooling Methods 0.000 claims description 9
- 230000007246 mechanism Effects 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 5
- 238000009499 grossing Methods 0.000 claims description 3
- 238000006116 polymerization reaction Methods 0.000 claims description 3
- 230000006798 recombination Effects 0.000 claims description 3
- 238000005215 recombination Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 2
- 208000024827 Alzheimer disease Diseases 0.000 abstract description 4
- 238000013399 early diagnosis Methods 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 39
- 238000002474 experimental method Methods 0.000 description 6
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 208000012902 Nervous system disease Diseases 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000000971 hippocampal effect Effects 0.000 description 1
- 238000011423 initialization method Methods 0.000 description 1
- 238000002370 liquid polymer infiltration Methods 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 230000004660 morphological change Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20172—Image enhancement details
- G06T2207/20192—Edge enhancement; Edge preservation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30016—Brain
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The invention discloses a hippocampus extraction method of a human brain nuclear magnetic resonance image based on a 3D neural network, which comprises the steps of preprocessing an original image data set and a label, constructing a 3D hippocampus segmentation neural network model, defining a boundary enhancement loss function, optimizing the boundary enhancement loss function, and detecting a preprocessed image to be detected by utilizing a trained 3D hippocampus segmentation neural network model to obtain a hippocampus extraction result. The invention utilizes the 3D hippocampus segmentation neural network model to realize the efficient, automatic and accurate segmentation of the hippocampus structure in the human brain MRI, and can reduce the time of doctors in early diagnosis of the Alzheimer disease.
Description
Technical Field
The invention belongs to the technical field of hippocampus segmentation, and particularly relates to a method for extracting a hippocampus of a human brain nuclear magnetic resonance image based on a 3D neural network.
Background
Hippocampus structure is an important tissue structure in the human brain, and morphological analysis thereof is important for detecting and diagnosing the clinical condition of the brain. The hippocampus structure is related to the memory mechanism, and the morphological change of the hippocampus structure is closely related to the Alzheimer disease and other nervous system diseases. The estimation of whether or not hippocampus is atrophied from Magnetic Resonance Images (MRI) is considered as one of the key techniques for diagnosing alzheimer's disease. However, manually segmenting the hippocampus structure in brain MRI is time consuming, labor intensive, and prone to error due to factors such as the small size of the hippocampus in the brain, complex morphology, and lack of sharp boundaries with surrounding structures. Therefore, how to rapidly divide the structure of the hippocampus in nuclear magnetic resonance becomes a leading-edge research topic in recent years at home and abroad, and the method not only has important academic value, but also has important social significance and wide application prospect.
The existing hippocampus segmentation methods are divided into manual segmentation, traditional segmentation methods and deep learning-based methods. Manual segmentation is susceptible to factors that are considered to make the segmentation effect unstable. The traditional segmentation method also consumes too long time and has inaccurate segmentation effect.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a method for extracting the hippocampus of the human brain nuclear magnetic resonance image based on a 3D neural network, aiming at solving the problems of overlong time consumption and low precision of the automatic segmentation technology of the hippocampus in the human brain nuclear magnetic resonance image.
In order to achieve the purpose of the invention, the invention adopts the technical scheme that:
a method for extracting hippocampus based on a human brain nuclear magnetic resonance image of a 3D neural network comprises the following steps:
s1, acquiring an original image data set of a human brain nuclear magnetic resonance image containing a 3D label as a training set, and preprocessing the original image data set and the label;
s2, constructing a 3D hippocampus segmentation neural network model consisting of a coding part and a channel attention decoding part, extracting characteristics of the image preprocessed in the step S1 through the coding part, and performing polymerization recombination on the characteristics through the decoding part containing a channel attention mechanism to finally obtain a hippocampus segmentation probability graph;
s3, extracting a boundary from the training set label data preprocessed in the step S1 by using the set convolutional layer, and defining a boundary enhancement loss function;
s4, optimizing the boundary enhancement loss function defined in the step S3 by using a back propagation algorithm, and training the 3D hippocampus segmentation neural network model constructed in the step S2;
and S5, detecting the image to be detected after being preprocessed in the step S1 by using the 3D hippocampus segmentation neural network model trained in the step S4 to obtain a hippocampus segmentation probability map, setting a threshold value to judge the category of each point in the hippocampus segmentation probability map, and further obtaining a hippocampus segmentation result.
Further, the preprocessing of the original image data set in step S1 specifically includes the following sub-steps:
s11, performing center region clipping on the original image data and the label image data corresponding to the original image data to obtain an image set with the size of H x W x D after clipping;
and S12, carrying out normalization processing on the single image on the cut image set obtained in the step S11, namely counting the mean value and the standard deviation of the brightness values of each group of image data, and dividing the mean value of the brightness values of the single image by the standard deviation to obtain a normalization result.
Further, the 3D hippocampus segmentation neural network model composed of the encoding portion and the channel attention decoding portion and constructed in step S2 specifically includes a convolution module, a residual module and a channel attention module, where the convolution module is composed of convolution, group regularization and ReLU activation, the residual module is composed of convolution, group regularization, ReLU activation and point-to-point addition sum operation, and the channel attention module is composed of convolution, group regularization, ReLU activation, sigmoid activation, point-to-point addition sum operation and point-to-mul operation.
Further, the processing procedure of the 3D hippocampus segmentation neural network model in step S2 is specifically divided into seven sub-steps, the first sub-step includes a convolution module, a residual module and a maximum pooling operation, the second sub-step includes two residual modules and a maximum pooling operation, the third sub-step includes two residual modules and a maximum pooling operation, the fourth sub-step includes two convolution modules and an upsampling operation, the fifth sub-step includes a convolution module, a channel attention module, a residual module and an upsampling operation, the sixth sub-step includes a convolution module, a channel attention module, a residual module and an upsampling operation, and the seventh substep comprises a convolution module, a channel attention module, a residual module and a convolution operation.
Further, the step S3 specifically includes the following sub-steps:
s31, setting a convolution layer, initializing a convolution kernel by using a sobel operator, and performing convolution operation on label data corresponding to the training image input into the network model constructed in the step S2 to extract a boundary;
s32, defining a boundary enhancement logarithmic dice loss function, a boundary enhancement cross entropy loss function and a definition type precision loss function by using the boundary extracted in the step S31;
and S33, setting the weight of each loss function, and adding to obtain a boundary enhancement loss function.
Further, the boundary enhancement loss function is specifically expressed as:
LBE=ωBEDLBED+ωBECLBEC+ωPLP
wherein L isBEDEnhancing the logarithmic dice loss function for the boundary, LBECEnhancing the cross-entropy loss function for the boundary, LPFor the precision-like loss function,LBEC=∑x-αB(x)(ln(PH(x))δH(x)+ln(PBG(x))δBG(x)),x is the position of the voxel point, PH(x) Predicting x points as a confidence value, P, of the hippocampus for the network modelBG(x) Predicting the x point as a confidence value of the hippocampus for the network model; delta when truth label of point x is hippocampusH(x) Is 1, otherwise is 0;delta when the truth label of point x is backgroundBG(x) Is 1, otherwise is 0; α when x point is considered as boundaryB(x) Is α, otherwise is 1; ε is the smoothing coefficient, ωBED、ωBECAnd ωPAre respectively LBED、LBECAnd LPThe weight of (c).
Further, the step S4 of optimizing the boundary enhancement loss function defined in the step S3 by using a back propagation algorithm specifically includes:
randomly initializing the 3D hippocampus segmentation neural network model constructed in the step S2, inputting the image preprocessed in the step S1 into the network model to obtain a hippocampus segmentation probability map, calculating a loss value by using the obtained hippocampus segmentation probability map and an original label corresponding to the input image through a boundary enhancement loss function defined in the step S3, and selecting a model weight which enables the model loss to be minimum in a training round.
The invention has the following beneficial effects:
(1) the invention provides a three-dimensional full-convolution U-shaped network structure with an attention mechanism, which can efficiently multiplex low-level spatial structure information to enhance the discrimination capability of the network on the difference between classes;
(2) the invention designs a new boundary-enhanced loss function, which can balance the class imbalance problem caused by the undersize structure of the hippocampus and make the network model more sensitive to the boundary of the hippocampus structure, and trains a network model with higher segmentation precision and more robust than other loss functions by using the loss function;
(3) the invention utilizes the 3D hippocampus segmentation neural network model to realize the efficient, automatic and accurate segmentation of the hippocampus structure in the human brain MRI, and can reduce the time of doctors in early diagnosis of the Alzheimer disease.
Drawings
FIG. 1 is a schematic flow chart of a method for extracting hippocampus based on a 3D neural network nuclear magnetic resonance image of a human brain according to the present invention;
FIG. 2 is a schematic diagram of a 3D hippocampus segmentation neural network model structure according to the present invention;
FIG. 3 is a schematic diagram of the combination of the modules in the 3D hippocampus segmentation neural network model according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Referring to fig. 1, an embodiment of the present invention provides a method for extracting hippocampus of a human brain nmr image based on a 3D neural network, including the following steps S1 to S5:
s1, acquiring an original image data set of a human brain nuclear magnetic resonance image containing a 3D label as a training set, and preprocessing the original image data set and the label;
in this example, the acquired raw image data set contains 130 sets of NIFTI formatted brain MRI hippocampal image files with a size of 197 × 233 × 189.
The method comprises the following steps of preprocessing the acquired original image data set:
s11, performing center region clipping on the original image data and the label image data corresponding to the original image data to obtain an image set with the size of H x W x D after clipping;
in this embodiment, the center point region of the 130 groups of original images and the corresponding label images are clipped, so as to obtain an image set with a size of 64 × 96 after clipping.
And S12, carrying out normalization processing on the single nuclear magnetic resonance image set obtained in the step S11 after clipping, namely counting the mean value and the standard deviation of the brightness values of each group of image data, and dividing the mean value of the brightness values of the single image by the standard deviation to obtain a normalization result.
S2, constructing a 3D hippocampus segmentation neural network model consisting of a coding part and a channel attention decoding part, extracting characteristics of the image preprocessed in the step S1 through the coding part, and performing polymerization recombination on the characteristics through the decoding part containing a channel attention mechanism to finally obtain a hippocampus segmentation probability graph;
in this embodiment, the present invention provides a 3D hippocampus segmentation full convolution neural network model with attention mechanism based on boundary enhancement loss to solve the problem of class imbalance caused by an excessively small hippocampus structure in a nuclear magnetic resonance image and the problem of small inter-class difference caused by fuzzy boundaries between the hippocampus structure and its surrounding tissues.
Referring to fig. 2 and 3, the 3D hippocampus segmentation neural network model includes a coding portion and a Channel Attention decoding portion, and the coding and decoding of the model specifically includes a Convolution Block (CB), a Residual Block (RB), and a Channel Attention Block (CAB); FIG. 3(a) is a convolution module consisting of convolution, group regularization, ReLU activation; FIG. 3(b) is a residual block, consisting of a 3D channel weighting block and sum, group regularization, ReLU activation; fig. 3(c) shows that the channel attention module is composed of convolution, group regularization, ReLU activation, sigmoid activation, point-by-point addition sum operation, and point-by-mul operation.
The 3D hippocampus segmentation neural network model extracts the characteristics of the image preprocessed in the step S1 through a coding part, and then segments the characteristics through a decoding part with a channel attention mechanism, and only one image is sent to network training each time.
Referring to fig. 2, the processing process of the 3D hippocampus segmentation neural network model is divided into seven sub-steps; the first substep comprises a convolution module, a residual error module and a 2 multiplied by 2 maximum pooling operation with the step length of 2, the number of convolution kernels in the convolution module and the residual error module is 64, the group normalization adopted in the modules is to divide every two channels into one group for normalization, and the group normalization mentioned in the invention adopts the same strategy; the second substep consists of two residual modules and a 2 × 2 × 2 maximal pooling operation with step size of 2, where the number of convolutional kernels per convolutional layer is 64; the third substep consists of two residual modules and a 2 × 2 × 2 max pooling operation with step size of 2, where the number of convolutional kernels per convolutional layer is 128; the fourth substep comprises two convolution modules and an upsampling operation with a sampling rate of 2, wherein the number of convolution kernels of each convolution layer in the three modules is respectively 256, 128 and 128; the fifth substep consists of a convolution module, a channel attention module, a residual module and an upsampling operation with a sampling rate of 2, wherein the number of convolution kernels per convolution layer is 128; the sixth substep comprises a convolution module, a channel attention module, a residual module and an upsampling operation with a sampling rate of 2, wherein the number of convolution kernels of each convolution layer is 64; the seventh substep includes a convolution operation with 2 convolution kernels, 3 × 3 × 3 convolution kernels, a channel attention module, a residual module, and a convolution kernel, where the number of convolution kernels in the convolution module and the channel attention module is 64, the number of convolution kernels in the residual module is 32, and all convolution steps are 1 in this embodiment. And adding softmax operation after the last convolution layer to obtain a final hippocampus segmentation probability map.
S3, extracting the Boundary of the training set label data preprocessed in step S1 by using the set convolutional layer, and defining a Boundary enhancement Loss function (BEL).
In this embodiment, step S3 specifically includes the following sub-steps:
s31, setting a convolution layer, initializing a convolution kernel by using a sobel operator, and performing convolution operation on label data corresponding to the training image input into the network model constructed in the step S2 to extract a boundary;
the single image is propagated in the forward direction to obtain the result of a single iteration, and the labels are firstly respectively fed into two convolution layers with fixed weights to determine the boundaries of the labels. Wherein, the weights of the two convolution layers are respectively initialized to two convolution kernels G of the Sobel operatorxAnd GyAnd performing convolution operation on the three-dimensional label data contained in the training set by taking the sagittal plane as mini-batch to extract a boundary.
Marking an initial label picture as A, obtaining the gradient size G of a sagittal plane,
the tag picture a is a hippocampus region with a luminance value of 1 and a background region with a value of 0. And setting a boundary judgment threshold alpha according to the obtained sagittal plane gradient G, and determining the point of G which is greater than the threshold alpha as the boundary of the hippocampus structure.
S32, defining a boundary enhancement logarithmic dice loss function, a boundary enhancement cross entropy loss function and a definition type precision loss function by using the boundary extracted in the step S31;
wherein, a boundary enhancement logarithm dice loss function L is definedBEDComprises the following steps:
defining a boundary-enhanced cross-entropy loss function LBECComprises the following steps:
LBEC=∑x-αB(x)(ln(PH(x))δH(x)+ln(PBG(x))δBG(x))
defining class precision loss function LPIs composed of
Wherein x is the position of a voxel point, PH(x) Predicting the x-point as a confidence value, P, of the hippocampus for the network modelBG(x) Predicting the x point as a confidence value of the hippocampus for the network model; delta when the truth label of the x point is hippocampusH(x) Is 1, otherwise is 0; delta when the truth label of point x is backgroundBG(x) Is 1, otherwise is 0; α when x point is considered as boundaryB(x) Is α, otherwise is 1; epsilon is a smoothing coefficient, and epsilon is 0.0001 to prevent denominator from being 0.
S33, setting the weight of each loss function, adding to obtain the boundary enhancement loss function LBEExpressed as:
LBE=ωBEDLBED+ωBECLBEC+ωPLP
wherein, ω isBED、ωBECAnd ωPAre respectively LBED、LBECAnd LPThe weight of (c).
S4, optimizing the boundary enhancement loss function defined in the step S3 by using a back propagation algorithm, and training the 3D hippocampus segmentation neural network model constructed in the step S2.
In this embodiment, the boundary enhancement loss function defined in the back propagation algorithm optimization step S3 is specifically:
randomly initializing the 3D hippocampus segmentation neural network model constructed in the step S2, inputting the image preprocessed in the step S1 into the network model to obtain a hippocampus segmentation probability map, calculating a loss value by using the obtained hippocampus segmentation probability map and an original label corresponding to the input image through a boundary enhancement loss function defined in the step S3, and selecting a model weight which enables the model loss to be minimum in a training round.
And initializing the 3D hippocampus segmentation neural network model according to the model weight obtained by optimizing the boundary enhancement loss function.
And S5, detecting the image to be detected after being preprocessed in the step S1 by using the 3D hippocampus segmentation neural network model trained in the step S4 to obtain a hippocampus segmentation probability map, setting a threshold to judge the category of each point in the probability map, and further obtaining a hippocampus segmentation result.
In this embodiment, the image to be detected preprocessed in step S1 is input into the 3D hippocampus segmentation neural network model trained in step S4, and a threshold is set to 0.5, so that the final segmentation result is the hippocampus extraction result.
To verify the effectiveness of the method of the present invention, experiments were performed on the EADC-ADNI database, and the experimental data of the present invention consisted of 130 brain MRI images, with 130 groups including real patients and healthy comparison population. To validate the performance of the model, the data were divided into 5 pieces, 5-fold cross-validation experiments were used, 4 pieces for training and 1 piece for testing until all data were tested. Regarding the optimization algorithm of the model, an Adam algorithm is adopted, the initial learning rate is set to be 0.001, and the weight initialization uses an Xavier uniform distribution initialization method.
The hardware equipment is as follows: the Intel Core i7-9700K CPU @4.2GHz of the processor; memory (RAM)24.0 GB; an independent graphics card, NVIDIA GeForce GTX 1080 Ti; system type, Ubuntu 16.04; development tools, Python and Keras frame based on the rear end of the tensierflow.
The Dice similarity is taken as an evaluation index of the experiment and is expressed as
Wherein TP represents the total number of voxels in the image for which the true value is hippocampus and which are predicted to hippocampus, FN represents the total number of voxels in the image for which the true value is hippocampus but the prediction result is background, and FP represents the total number of voxels in the image for which the true value is background but the prediction result is hippocampus.
The embodiment verification comprises the following three parts. The first part respectively trains different network structures by using the loss function BEL provided by the invention to verify the effectiveness of the provided network structures; 2) training the network structure proposed by the invention by using different loss functions to verify the effectiveness of the proposed loss functions; 3) finally, the method provided by the invention is compared with two groups of algorithms which are the same as the two algorithms for segmenting the medical image by utilizing the three-dimensional convolutional neural network.
1. Loss function contrast
The convolutional neural network designed by the invention is trained by using a Loss function designed by the invention and small Batch Weighted Cross Entropy (BWCE) Loss and Exponential logarithmic Loss (EL) respectively. Wherein, ω is the loss function of the present inventionBED、ωBEC、ωPRespectively 4, 5 and 1, the best effect is achieved. Statistical loss function the results of the experiments are shown in table 1. From the Dice column of Table 1, it can be seen that the loss function BEL proposed by the present invention obtains the highest Dice precision, and from the convert Epoch column, it can be seen that it exhibits the fastestAnd (4) convergence speed.
TABLE 1 segmentation results (mean + -std%)
2. Network structure comparison
The set of experiments respectively train two different 3D convolutional neural networks with the loss function proposed by the present invention, and compare the network structures proposed by the present invention, and the statistical comparison experiment results are shown in Table 2.
TABLE 2 segmentation results (mean + std%) from training different networks using the proposed loss function
3. Compared with the prior method
Table 3 statistics of the performance of the various methods on the same dataset and their proposed years. Wherein FreeScherfer and Label Fusion use different cases in the same dataset to obtain results. Other methods were performed using the same case. It can be seen from the Dice column of table 3 that the present invention is superior to other methods.
TABLE 3 comparison of Performance of the method of the present invention with other algorithms (mean + -std%)
It will be appreciated by those of ordinary skill in the art that the embodiments described herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited embodiments and examples. Those skilled in the art can make various other specific changes and combinations based on the teachings of the present invention without departing from the spirit of the invention, and these changes and combinations are within the scope of the invention.
Claims (6)
1. A hippocampus extraction method of a human brain nuclear magnetic resonance image based on a 3D neural network is characterized by comprising the following steps:
s1, acquiring an original image data set of a human brain nuclear magnetic resonance image containing a 3D label as a training set, and preprocessing the original image data set and the label;
s2, constructing a 3D hippocampus segmentation neural network model consisting of a coding part and a channel attention decoding part, extracting characteristics of the image preprocessed in the step S1 through the coding part, and performing polymerization recombination on the characteristics through the decoding part containing a channel attention mechanism to finally obtain a hippocampus segmentation probability map;
s3, extracting a boundary from the training set label data preprocessed in the step S1 by using the set convolutional layer, and defining a boundary enhancement loss function; the boundary enhancement loss function is specifically expressed as:
LBE=ωBEDLBED+ωBECLBEC+ωPLP
wherein L isBEDEnhancing the logarithmic dice loss function for the boundary, LBECEnhancing the cross-entropy loss function for the boundary, LPFor the precision-like loss function, x is the position of the voxel point, PH(x) Predicting the x-point as a confidence value, P, of the hippocampus for the network modelBG(x) Predicting the x point as a confidence value of the hippocampus for the network model; delta when the truth label of the x point is hippocampusH(x) Is 1, otherwise is 0; delta when the truth label of point x is backgroundBG(x) Is 1, otherwise is 0; α when x point is considered as boundaryB(x) Is α, otherwise is 1; ε is the smoothing coefficient, ωBED、ωBECAnd ωPAre respectively LBED、LBECAnd LPThe weight of (c);
s4, optimizing the boundary enhancement loss function defined in the step S3 by using a back propagation algorithm, and training the 3D hippocampus segmentation neural network model constructed in the step S2;
and S5, detecting the image to be detected after being preprocessed in the step S1 by using the 3D hippocampus segmentation neural network model trained in the step S4 to obtain a hippocampus segmentation probability map, setting a threshold value to judge the category of each point in the hippocampus segmentation probability map, and further obtaining a hippocampus segmentation result.
2. The method for extracting hippocampus of human brain nmr image based on 3D neural network of claim 1, wherein the preprocessing of the original image dataset in step S1 includes the following sub-steps:
s11, performing center region clipping on the original image data and the label image data corresponding to the original image data to obtain an image set with the size of H x W x D after clipping;
and S12, carrying out normalization processing on the single image on the cut image set obtained in the step S11, namely counting the mean value and the standard deviation of the brightness values of each group of image data, and dividing the mean value of the brightness values of the single image by the standard deviation to obtain a normalization result.
3. The method for extracting hippocampus of human brain NMR image based on 3D neural network as claimed in claim 1, wherein the 3D hippocampus segmentation neural network model composed of an encoding part and a channel attention decoding part constructed in step S2 includes a convolution module, a residual module and a channel attention module, the convolution module is composed of convolution, group regularization and ReLU activation, the residual module is composed of convolution, group regularization, ReLU activation and point-by-point addition sum operation, and the channel attention module is composed of convolution, group regularization, ReLU activation, sigmoid activation, point-by-point addition sum operation and point-by-mul operation.
4. The method for extracting hippocampus of human brain NMR image based on 3D neural network of claim 2, wherein the process of the 3D hippocampus segmentation neural network model in step S2 is divided into seven sub-steps, the first sub-step includes a convolution module, a residual error module and a max pooling operation, the second sub-step includes two residual error modules and a max pooling operation, the third sub-step includes two residual error modules and a max pooling operation, the fourth sub-step includes two convolution modules and an up-sampling operation, the fifth sub-step includes a convolution module, a channel attention module, a residual error module and an up-sampling operation, the sixth sub-step includes a convolution module, a channel attention module, a residual error module and an up-sampling operation, the seventh substep comprises a convolution module, a channel attention module, a residual module, and a convolution operation.
5. The method for extracting hippocampus of human brain NMR image based on 3D neural network of claim 1, wherein the step S3 includes the following sub-steps:
s31, setting a convolution layer, initializing a convolution kernel by using a sobel operator, and performing convolution operation on label data corresponding to the training image input into the network model constructed in the step S2 to extract a boundary;
s32, defining a boundary enhancement logarithmic dice loss function, a boundary enhancement cross entropy loss function and a definition type precision loss function by using the boundary extracted in the step S31;
and S33, setting the weight of each loss function, and adding to obtain a boundary enhancement loss function.
6. The method for extracting hippocampus of human brain nmr image based on 3D neural network of claim 1, wherein the step S4 of optimizing the boundary enhancement loss function defined in step S3 by using back propagation algorithm is specifically:
randomly initializing the 3D hippocampus segmentation neural network model constructed in the step S2, inputting the image preprocessed in the step S1 into the network model to obtain a hippocampus segmentation probability map, calculating a loss value by using the obtained hippocampus segmentation probability map and an original label corresponding to the input image through a boundary enhancement loss function defined in the step S3, and selecting a model weight which enables the model loss to be minimum in a training round.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911179566.4A CN110969626B (en) | 2019-11-27 | 2019-11-27 | Method for extracting hippocampus of human brain nuclear magnetic resonance image based on 3D neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911179566.4A CN110969626B (en) | 2019-11-27 | 2019-11-27 | Method for extracting hippocampus of human brain nuclear magnetic resonance image based on 3D neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110969626A CN110969626A (en) | 2020-04-07 |
CN110969626B true CN110969626B (en) | 2022-06-07 |
Family
ID=70031832
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201911179566.4A Active CN110969626B (en) | 2019-11-27 | 2019-11-27 | Method for extracting hippocampus of human brain nuclear magnetic resonance image based on 3D neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110969626B (en) |
Families Citing this family (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20210334975A1 (en) * | 2020-04-23 | 2021-10-28 | Nvidia Corporation | Image segmentation using one or more neural networks |
CN111696119A (en) * | 2020-06-02 | 2020-09-22 | 广州柏视医疗科技有限公司 | Nuclear magnetic resonance image hippocampus segmentation method and device |
CN111985409B (en) * | 2020-08-21 | 2021-03-16 | 四川省人工智能研究院(宜宾) | Method for real-time street scene segmentation |
CN112164074B (en) * | 2020-09-22 | 2021-08-10 | 江南大学 | 3D CT bed fast segmentation method based on deep learning |
CN112329867B (en) * | 2020-11-10 | 2021-05-25 | 宁波大学 | MRI image classification method based on task-driven hierarchical attention network |
CN112674720B (en) * | 2020-12-24 | 2022-03-22 | 四川大学 | Alzheimer disease pre-judgment method based on 3D convolutional neural network |
CN113052856A (en) * | 2021-03-12 | 2021-06-29 | 北京工业大学 | Hippocampus three-dimensional semantic network segmentation method based on multi-scale feature multi-path attention fusion mechanism |
US11813067B2 (en) * | 2021-06-01 | 2023-11-14 | National Cheng Kung University | System and method for Alzheimer's disease prediction using neural network, computer-readable recording medium with stored program, and computer program product |
CN113496496B (en) * | 2021-07-07 | 2023-04-07 | 中南大学 | MRI image hippocampus region segmentation method based on multiple losses and multiscale characteristics |
CN113870243A (en) * | 2021-10-11 | 2021-12-31 | 烟台大学 | Magnetic resonance brain image hippocampus segmentation method |
CN114533066B (en) * | 2022-04-28 | 2022-08-19 | 之江实验室 | Social anxiety assessment method and system based on composite expression processing brain network |
CN115311453B (en) * | 2022-06-27 | 2023-06-27 | 河南省人民医院 | Automatic sea horse segmentation method and system based on V-Net and combined loss function |
CN115345952B (en) * | 2022-08-10 | 2023-08-11 | 华中科技大学协和深圳医院 | Magnetic resonance image processing method and system based on neural network |
CN115240014B (en) * | 2022-09-21 | 2023-01-13 | 山东大学齐鲁医院 | Medical image classification system based on residual error neural network |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108922599A (en) * | 2018-06-27 | 2018-11-30 | 西南交通大学 | A kind of accurate mask method of medical image lesion point based on MIL |
CN109215035A (en) * | 2018-07-16 | 2019-01-15 | 江南大学 | A kind of brain MRI hippocampus three-dimensional dividing method based on deep learning |
CN109584244A (en) * | 2018-11-30 | 2019-04-05 | 安徽海浪智能技术有限公司 | A kind of hippocampus dividing method based on Sequence Learning |
CN110120033A (en) * | 2019-04-12 | 2019-08-13 | 天津大学 | Based on improved U-Net neural network three-dimensional brain tumor image partition method |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11024025B2 (en) * | 2018-03-07 | 2021-06-01 | University Of Virginia Patent Foundation | Automatic quantification of cardiac MRI for hypertrophic cardiomyopathy |
-
2019
- 2019-11-27 CN CN201911179566.4A patent/CN110969626B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108922599A (en) * | 2018-06-27 | 2018-11-30 | 西南交通大学 | A kind of accurate mask method of medical image lesion point based on MIL |
CN109215035A (en) * | 2018-07-16 | 2019-01-15 | 江南大学 | A kind of brain MRI hippocampus three-dimensional dividing method based on deep learning |
CN109584244A (en) * | 2018-11-30 | 2019-04-05 | 安徽海浪智能技术有限公司 | A kind of hippocampus dividing method based on Sequence Learning |
CN110120033A (en) * | 2019-04-12 | 2019-08-13 | 天津大学 | Based on improved U-Net neural network three-dimensional brain tumor image partition method |
Non-Patent Citations (6)
Title |
---|
AIHGF.医学图像分割之 Dice Loss.《https://www.aiuai.cn/aifarm1159.html》.2019,第1-7页. * |
Carole H. Sudre 等.Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations.《ML-CDS 2017: Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support》.2017,第10553卷第240–248页. * |
Ken C. L. Wong 等.3D Segmentation with Exponential Logarithmic Loss for Highly Unbalanced Object Sizes.《Medical Image Computing and Computer Assisted Intervention – MICCAI 2018》.2018,第11072卷第612–619页. * |
L. Wang 等.Nested Dilation Network (NDN) for Multi-Task Medical Image Segmentation.《IEEE Access》.2019,第7卷第44676-44685页. * |
曹平 等.结合序列学习和U型网络的海马体分割方法.《计算机辅助设计与图形学学报》.2019,第31卷(第8期),第1382-1390页. * |
颜宇 等.基于边界增强损失的海马体分割算法.《光电子•激光》.2020,第31卷(第3期),第299-309页. * |
Also Published As
Publication number | Publication date |
---|---|
CN110969626A (en) | 2020-04-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110969626B (en) | Method for extracting hippocampus of human brain nuclear magnetic resonance image based on 3D neural network | |
Soudani et al. | An image-based segmentation recommender using crowdsourcing and transfer learning for skin lesion extraction | |
Meng et al. | Liver tumor segmentation based on 3D convolutional neural network with dual scale | |
Ahirwar | Study of techniques used for medical image segmentation and computation of statistical test for region classification of brain MRI | |
CN109949276B (en) | Lymph node detection method for improving SegNet segmentation network | |
Mahapatra | Analyzing training information from random forests for improved image segmentation | |
WO2021203795A1 (en) | Pancreas ct automatic segmentation method based on saliency dense connection expansion convolutional network | |
CN112446891B (en) | Medical image segmentation method based on U-Net network brain glioma | |
Wen et al. | A novel statistical cerebrovascular segmentation algorithm with particle swarm optimization | |
CN112270666A (en) | Non-small cell lung cancer pathological section identification method based on deep convolutional neural network | |
CN107766874B (en) | Measuring method and measuring system for ultrasonic volume biological parameters | |
Mahapatra et al. | Visual saliency based active learning for prostate mri segmentation | |
CN109492668B (en) | MRI (magnetic resonance imaging) different-phase multimode image characterization method based on multi-channel convolutional neural network | |
CN109447963A (en) | A kind of method and device of brain phantom identification | |
CN113269799A (en) | Cervical cell segmentation method based on deep learning | |
CN116012353A (en) | Digital pathological tissue image recognition method based on graph convolution neural network | |
CN111738997A (en) | Method for calculating new coronary pneumonia lesion area ratio based on deep learning | |
CN112434172A (en) | Pathological image prognosis feature weight calculation method and system | |
CN117036386A (en) | Cervical MRI image self-supervision segmentation method for generating data by using diffusion model | |
CN115496720A (en) | Gastrointestinal cancer pathological image segmentation method based on ViT mechanism model and related equipment | |
CN110992309B (en) | Fundus image segmentation method based on deep information transfer network | |
CN117036288A (en) | Tumor subtype diagnosis method for full-slice pathological image | |
Feng et al. | Automatic localization and segmentation of focal cortical dysplasia in FLAIR‐negative patients using a convolutional neural network | |
CN109740669B (en) | Breast cancer pathological image classification method based on depth feature aggregation | |
Soleymanifard et al. | Segmentation of whole tumor using localized active contour and trained neural network in boundaries |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |